Metal material intelligent self-consistent screening method and system based on multi-modal database

By constructing a multimodal database and a deep learning network, extracting shared features, and establishing a multimodal coherent model, the problems of spectral distortion and fragmented determination of compositional properties in industrial scenarios of traditional atomic emission spectroscopy are solved. This enables intelligent self-consistent screening of the composition and properties of metallic materials, improving the accuracy and reliability of identification.

CN121009416APending Publication Date: 2025-11-25NCS TESTING TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202511134083.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-08-13
Publication Date
2025-11-25

AI Technical Summary

Technical Problem

Traditional atomic emission spectroscopy faces challenges in industrial settings, including complex spectral distortion, disconnect between component analysis and material property determination, underutilization of the collaborative potential of multimodal spectral data, and a lack of self-consistent verification mechanisms. These issues lead to decreased calibration accuracy and a high misjudgment rate.

Method used

A multimodal database is constructed, shared features are extracted through deep learning networks, a multimodal coherence model is established, cross-modal feature fusion is performed, and a self-consistent verification engine is designed to achieve synchronous optimization and bidirectional verification of composition quantification and material properties.

Benefits of technology

It improves the accuracy of metal material classification and identification, and significantly enhances the reliability and accuracy of detection in complex environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121009416A_ABST
    Figure CN121009416A_ABST
Patent Text Reader

Abstract

The invention discloses a metal material intelligent self-consistent screening method and system based on a multi-modal database, and the method comprises the steps: collecting atomic emission spectrum data of a metal material in different forms and surface states, and associating material attributes, collection parameters and environmental parameters to generate a multi-modal database; constructing a feature extraction network by adopting a deep learning network, performing cross-modal feature fusion on data in the multi-modal database, extracting shared features reflecting essential attributes of a material, training a plurality of sub-task models based on the shared features, and performing weighted joint optimization to obtain a multi-modal coherent model; and carrying out self-consistent verification on a component content prediction result and a material attribute classification result output by the model, and correcting the results. According to the method, the multi-modal database is constructed, the shared features are extracted, the multi-modal coherent model is established, component quantification and material attribute classification are synchronously optimized, self-consistent verification is carried out, and the accuracy of metal material classification and recognition is greatly improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of metal material classification and identification technology, and in particular to an intelligent self-consistent screening method and system for metal materials based on a multimodal database. Background Technology

[0002] Atomic emission spectroscopy (AES), a classic technique for analyzing the composition of materials, is based on the characteristic radiation produced by stimulated atomic transitions. It achieves qualitative and quantitative elemental analysis by resolving spectral wavelengths and intensities. Due to its high sensitivity, simultaneous detection of multiple elements, and the elimination of complex pretreatment, AES has become the mainstream technique for analyzing the composition of metallic materials. However, the complexity of real-world metallic materials in industrial settings poses significant challenges to traditional AES technology.

[0003] First, spectral distortion in industrial settings is more complex. Real-world metallic materials often exhibit diverse forms (such as alloys and composite coatings) and heterogeneous surface states (oxide layers and contaminant layers), leading to nonlinear distortion of spectral signals. Traditional AES relies on standard sample databases established under laboratory conditions, and their spectral characteristics cannot cover spectral distortion caused by dynamic factors such as temperature fluctuations, matrix effects, and surface contamination in industrial settings. This is especially true for highly alloyed or multiphase materials, where the correction accuracy decreases significantly.

[0004] Secondly, composition analysis and material property determination have long been separated. Existing technologies usually treat quantitative composition (such as element content) and material properties (such as mechanical properties and corrosion resistance) as independent tasks, lacking a cross-dimensional correlation analysis mechanism. For example, although the content of a certain element meets the standard, its distribution pattern or synergistic effect with other elements may directly affect the material properties. The traditional one-way judgment logic is prone to misjudgment.

[0005] Furthermore, the collaborative potential of multimodal spectral data has not been fully explored. Although current AES systems can collect multimodal information such as full-spectrum spectra, excitation energy parameters, and environmental variables, mainstream algorithms only extract local features (such as characteristic spectral lines of specific elements), ignoring the global correlation between spectra and the coupling effect of operating parameters on the excitation process. When there are small fluctuations in composition or abrupt changes in surface state, the quantitative and classification results are prone to significant deviations.

[0006] Furthermore, the lack of a self-consistent verification mechanism exacerbates the potential for unreliability in analysis. Existing technologies struggle to achieve logical closed-loop verification between component content and material properties. For instance, abnormal fluctuations in the content of a certain element may be caused by instrument noise or changes in the actual composition, but traditional methods cannot intelligently correct these deviations through the reverse constraints of material performance indicators, resulting in a persistently high false alarm rate.

[0007] In summary, traditional AES technology has significant bottlenecks in terms of dynamic operating condition adaptability, cross-dimensional correlation modeling capabilities, and multi-source data fusion analysis. There is an urgent need to establish a spectral database and intelligent verification mechanism that integrates actual operating condition characteristics. Summary of the Invention

[0008] The purpose of this invention is to provide an intelligent self-consistent screening method and system for metallic materials based on a multimodal database. The method involves constructing a multimodal database, extracting shared features, establishing a multimodal coherent model, simultaneously optimizing component quantification and material property classification, and designing a self-consistent verification engine to bidirectionally verify the material property-content logic, thereby greatly improving the accuracy of metallic material classification and identification.

[0009] To achieve the above objectives, the present invention provides the following solution:

[0010] A smart self-consistent screening method for metallic materials based on a multimodal database includes the following steps:

[0011] S1, Construct a multimodal database: Collect atomic emission spectral data of metallic materials under different morphologies and surface states, and associate them with material properties, acquisition parameters, and environmental parameters to generate a multimodal database; wherein, the morphology, surface state, acquisition parameters, and environmental parameters constitute multimodal metadata;

[0012] S2, Establish a multimodal coherent model: A feature extraction network is constructed using a deep learning network to perform cross-modal feature fusion on data in the multimodal database, extract shared features that reflect the essential properties of materials, train multiple sub-task models based on the shared features, and obtain a multimodal coherent model through weighted joint optimization. The multiple sub-task models include a material composition content prediction model, a material property classification model, and a multimodal metadata prediction model.

[0013] S3, Self-consistent classification verification: Perform self-consistent verification and correct the results of the model output component content prediction results and material property classification results. The self-consistent verification methods include material property-content cross-verification, multimodal consistency check, and historical data backtesting evaluation.

[0014] Furthermore, in S1, different forms include powder, block, thin film, and coating; different surface states include oxide layer, corrosion layer, and polished surface; atomic emission spectral data include full-spectrum wavelength, intensity, signal-to-noise ratio, and spectral line broadening parameters; material properties include metal grade, chemical composition content, heat treatment state, and processing technology; acquisition parameters include laser energy, laser pulse width, laser wavelength, number of ablation cycles, detection height, detection angle, and spot position; environmental parameters include temperature and air pressure.

[0015] Furthermore, the multimodal database also includes: annotation information and knowledge graph, wherein the annotation information includes manually verified material property-content correspondence, abnormal spectrum markers, reference spectrum calibration, and cross-modal data association tags, and the knowledge graph includes material multimodal mapping relationships.

[0016] Furthermore, the multimodal database supports the following functions:

[0017] Dynamic updates: By incorporating newly acquired multimodal metadata and validation results through an incremental learning mechanism, the material properties - content statistical distribution are automatically updated;

[0018] Cross-modal retrieval: Quickly locate similar spectra based on morphology, surface condition, or acquisition parameters, and return historical data on corresponding material properties and content;

[0019] Data traceability: Record the acquisition equipment, operators, and environmental conditions for each spectrum to ensure the results are reproducible;

[0020] Semantic query: Supports retrieving relevant spectral data and model training results by material properties or component content range.

[0021] Furthermore, in S2, cross-modal feature fusion specifically includes:

[0022] Data augmentation: Align and register atomic emission spectral data of the same metallic material under different modal metadata to construct intermodal correlations;

[0023] Feature decoupling: Extracting shared features that reflect the essential properties of materials through a shared feature layer design;

[0024] Dynamic weight allocation: Adaptively adjusts the contribution of different modal metadata to feature extraction based on morphology or surface state.

[0025] Furthermore, in S2, the training method includes:

[0026] Independent training mode: The feature extraction network is optimized through unsupervised or self-supervised learning, and then the shared features obtained from pre-training are input into multiple sub-task models;

[0027] Joint training mode: The feature extraction network, content prediction branch, and material property classification branch are constructed into an end-to-end model, which is simultaneously optimized through a multi-task loss function;

[0028] Progressive training mode: First train the feature extraction network and a single subtask model, then introduce another subtask model for fine-tuning.

[0029] Furthermore, in S3, the material property-content cross-verification is: based on the standard component range corresponding to the material property, verifying the rationality of the component content prediction results;

[0030] Multimodal consistency check involves comparing the prediction results of the same metallic material under different acquisition parameters to eliminate random errors.

[0031] Historical data retrospective evaluation involves calling historical data under the same modality metadata configuration in the multimodal database to conduct confidence assessment.

[0032] Furthermore, the material property-content cross-verification includes:

[0033] Material property-driven validation: Based on the material property classification results output by the model, the standard content range corresponding to the material property is extracted from the multimodal database. If the predicted content of the component exceeds the range, the content correction is triggered.

[0034] Material content-driven verification: Based on the composition content prediction results, the matching material properties in the multimodal database are retrieved in reverse. If the current material property is not in the set, manual verification is initiated.

[0035] Dynamic threshold adjustment: Based on historical deviation data of similar metallic materials in a multimodal database, the material property-content compatibility threshold is adaptively adjusted.

[0036] On the other hand, the present invention also provides an intelligent self-consistent screening system for metallic materials based on a multimodal coherence model, applied to the above-mentioned intelligent self-consistent screening method for metallic materials based on a multimodal database, comprising:

[0037] The multimodal database construction module collects atomic emission spectrum data of metallic materials under different morphologies and surface states, and associates them with material properties, acquisition parameters, and environmental parameters to generate a multimodal database; wherein, the morphology, surface state, acquisition parameters, and environmental parameters constitute multimodal metadata;

[0038] The multimodal coherent model building module is used to construct a feature extraction network using a deep learning network, perform cross-modal feature fusion on data in a multimodal database, extract shared features that reflect the essential properties of materials, train multiple sub-task models based on the shared features, and obtain a multimodal coherent model through weighted joint optimization. The multiple sub-task models include a material composition content prediction model, a material property classification model, and a multimodal metadata prediction model.

[0039] The self-consistent classification verification module is used to perform self-consistent verification and correct the results of the model output component content prediction results and material property classification results. The self-consistent verification methods include material property-content cross-verification, multimodal consistency check, and historical data backtesting evaluation.

[0040] As can be seen from the above technical solutions, compared with the prior art, the intelligent self-consistent screening method and system for metallic materials based on a multimodal database provided by the present invention has the following beneficial effects:

[0041] In this method, a multimodal database stores multimodal metadata of actual materials and corresponding atomic emission spectrum data. During model training, shared feature representations reflecting the essential properties of materials are extracted from the multimodal database. Self-consistent classification verifies the output of component content prediction results and material property classification results, and corrects the results through self-consistent verification. This invention integrates multimodal database construction, cross-modal feature fusion, multi-sub-task model joint training, and material property-content mutual verification into a closed-loop feedback scheme. It utilizes the constructed multimodal database for shared feature extraction and trains the material component content prediction model and material property classification model synchronously or stepwise based on the extracted shared features. Finally, it achieves intelligent identification of metallic material composition and material properties through a self-consistent verification mechanism and dynamically updates the multimodal database, greatly improving the accuracy of metallic material classification and identification and significantly enhancing the detection reliability of metallic material classification and identification in complex environments. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart illustrating the intelligent self-consistent screening method for metallic materials based on a multimodal database according to the present invention.

[0044] Figure 2 This is a schematic diagram of the network architecture of a multimodal coherent model provided in an embodiment of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] The purpose of this invention is to provide an intelligent self-consistent screening method and system for metallic materials based on a multimodal database. The method involves constructing a multimodal database, extracting shared features, establishing a multimodal coherent model, simultaneously optimizing component quantification and material property classification, and designing a self-consistent verification engine to bidirectionally verify the material property-content logic, thereby greatly improving the accuracy of metallic material classification and identification.

[0047] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] like Figure 1 As shown, the intelligent self-consistent screening method for metallic materials based on a multimodal database provided by this invention includes the following steps:

[0049] S1, Construct a multimodal database:

[0050] Atomic emission spectral data (including full-spectrum wavelength, intensity, signal-to-noise ratio, and spectral line broadening parameters) of metallic materials in different forms (such as powder, block, thin film, coating, etc.) and surface states (such as oxide layer, corrosion layer, polished surface, etc.) are collected and correlated with material properties, acquisition parameters, and environmental parameters to generate a multimodal database.

[0051] The material properties include metal grade, chemical composition, heat treatment state (annealing, quenching, etc.), and processing technology (rolling, casting, etc.); the acquisition parameters include laser excitation parameters (laser energy, laser pulse width, laser wavelength, number of ablation cycles, etc.) and acquisition geometric parameters (including detection height, detection angle, and spot position, etc.); the environmental parameters include temperature and air pressure, etc.

[0052] The morphology, surface condition, acquisition parameters, and environmental parameters constitute multimodal metadata.

[0053] The multimodal database also includes: annotation information and knowledge graph. The annotation information includes manually verified material property-content correspondence, abnormal spectrum markers, reference spectrum calibration, and cross-modal data association tags. The knowledge graph includes material multimodal mapping relationships (mapping relationships between spectrum, property, and multimodal metadata) to support deep coupling and association of multiple material properties. The multimodal database provides cross-modal aligned training samples and associated labels for multimodal coherent models.

[0054] The multimodal database supports the following functions to enhance the application of material properties and content:

[0055] a) Dynamic update: By incorporating newly collected multimodal metadata and validation results through an incremental learning mechanism, the material properties - content statistical distribution are automatically updated;

[0056] b) Cross-modal retrieval: Quickly locate similar spectra based on material morphology, surface condition, or collected parameters, and return historical data on corresponding material properties and content;

[0057] c) Data traceability: Record the acquisition equipment, operators, and environmental conditions for each spectrum to ensure that the results are reproducible;

[0058] d) Semantic query: Supports retrieving relevant spectral data and model training results by material properties or component content range.

[0059] S2, Establish a multimodal coherent model:

[0060] A feature extraction network is constructed using a deep learning network to perform cross-modal feature fusion on data in a multimodal database, extracting shared features that reflect the essential properties of materials. Based on the shared features, multiple sub-task models are trained and then jointly optimized by weights to obtain a multimodal coherent model. The multiple sub-task models include a material composition content prediction model, a material property classification model, and a multimodal metadata prediction model.

[0061] The multimodal metadata prediction model can be divided into several sub-task models for predicting material morphology, surface condition, collected parameters, and environmental parameters.

[0062] The weighted joint optimization involves adjusting the weight coefficients of each sub-task model based on the gradient change of the task training loss of each sub-task model, so that the training rates of multiple sub-task models are within the same numerical range, until the overall loss function of multiple sub-task models meets the training cutoff condition.

[0063] Cross-modal feature fusion is achieved through the following methods:

[0064] a) Data augmentation: Align and register atomic emission spectrum data of the same metallic material under different modal metadata to construct intermodal correlations;

[0065] b) Feature Decoupling: A shared feature layer is designed to extract shared features reflecting the essential properties of materials. This invention constructs a spectral input layer, a feature extraction layer, and a shared feature layer using a deep learning network. The shared feature layer aims to extract features characterizing the essential properties of materials (such as equipment acquisition parameters, environmental parameters, and inherent material properties) through multimodal joint training, while shielding non-essential interference factors. The shared feature layer provides a unified feature representation for subsequent sub-tasks (such as composition prediction and property classification), supporting transfer learning.

[0066] c) Dynamic weight allocation: Adaptively adjust the contribution of different modal metadata to feature extraction based on material morphology or surface condition. For example, for surface oxidized materials, the weight of oxidation-related features in spectral data can be reduced, and material properties or environmental parameters can be used to assist feature fusion.

[0067] In S2, the training method includes:

[0068] Independent training mode: The feature extraction network (e.g., fully connected network FNN) is optimized through unsupervised or self-supervised learning, and the pre-trained shared features are then input into multiple sub-task models, such as material composition content prediction model and material property classification model.

[0069] Joint training mode: The feature extraction network, content prediction branch, and material property classification branch are constructed into an end-to-end model, which is simultaneously optimized through a multi-task loss function;

[0070] Progressive training mode: First train the feature extraction network and a single subtask model, then introduce a branch of another subtask model for fine-tuning.

[0071] The trained multimodal coherent model includes an input layer, a feature extraction layer, a shared feature layer, and multiple output layers, such as... Figure 2 As shown.

[0072] The system comprises the following layers: Input layer: receives atomic emission spectra; Feature extraction layer: progressively extracts multimodal specific features through a deep learning network; Shared feature layer: fuses cross-modal information to form a unified feature representation characterizing the essential properties of the material, learning the impact of equipment and environment on test results, implicitly defining the "capability boundary" of the current test conditions; Output layer (dedicated to sub-tasks): material composition output layer (e.g., elemental content prediction), material property classification output layer (e.g., metal grade classification), and multimodal metadata output layer (e.g., predicting geometry, surface condition, and acquired parameters).

[0073] When training a new subtask, transfer learning can be performed on the new subtask based on the shared features. During training, the input layer, feature extraction layer, and shared feature layer are locked, and only the output layer of the subtask needs to be trained, which greatly reduces the training data requirements and computational costs of the new task and improves the reusability of the model.

[0074] S3, Self-consistent classification verification: Perform self-consistent verification and correct the results of the model output component content prediction results and material property classification results. The self-consistent verification methods include material property-content cross-verification, multimodal consistency check, and historical data backtesting evaluation.

[0075] Among them, material property-content cross-verification is to verify the rationality of the component content prediction results based on the standard component range corresponding to the material properties; multimodal consistency check is to compare the prediction results of the same metal material under different acquisition parameters and eliminate random errors; historical data backtracking evaluation is to call historical data under the same modal metadata configuration in the multimodal database to evaluate the confidence level.

[0076] Specifically, the material property-content cross-verification includes:

[0077] Material property-driven validation: Based on the material property classification results output by the model, the standard content range corresponding to the material property is extracted from the multimodal database. If the predicted content of the component exceeds the range, the content correction is triggered.

[0078] Material content-driven verification: Based on the composition content prediction results, the matching material properties in the multimodal database are retrieved in reverse. If the current material property is not in the set, manual verification is initiated.

[0079] Dynamic threshold adjustment: Based on historical deviation data of similar metallic materials in a multimodal database, the material property-content compatibility threshold is adaptively adjusted.

[0080] The intelligent self-consistent screening method for metallic materials based on a multimodal database provided by this invention can be extended to the following scenarios, where material properties and content are the core analytical targets:

[0081] a) Online monitoring: Integrated into the industrial production line, it provides real-time feedback on material quality and triggers a self-consistent verification process, automatically comparing whether the current material properties and content meet the preset standards;

[0082] b) Environmental monitoring: For heavy metal particulate matter in soil and water, pollution source tracing is achieved through multimodal data collection (such as different particle sizes and oxidation states) combined with material property-content cross-verification.

[0083] c) Geological exploration: Multimodal LIBS analysis of core samples is performed to identify mineral types and invert mineralization processes through material properties;

[0084] d) Cultural Heritage Conservation: Analyze the multimodal spectrum of the corrosion layer of metallic artifacts and assess their preservation status by combining the evolution of material properties and content.

[0085] e) Aerospace: In-situ testing of high-temperature alloy components, overcoming high-temperature oxidation interference through multimodal data fusion, and achieving accurate identification of material properties and micro-area compositional segregation;

[0086] f) Nuclear industry: Remote LIBS testing of radioactive metallic materials, enhancing signal stability through multimodal data, and verifying material integrity by combining material property-content database.

[0087] Example 1

[0088] The intelligent self-consistent screening method for metallic materials based on a multimodal database provided in Specific Embodiment 1 of the present invention includes the following steps:

[0089] I. Constructing a multimodal database:

[0090] Spectral data were acquired through collaborative acquisition by two instruments:

[0091] LIBS: Acquire atomic emission spectroscopic data of metal samples with different morphologies, surface states, acquisition conditions, and acquisition geometric parameters.

[0092] ICP-OES: Provides true values ​​for element content.

[0093] By integrating over 10,000 sets of data and data annotations, a multimodal database of "spectral-material properties-multimodal metadata" was established.

[0094] II. Establishing a multimodal coherence model:

[0095] The training set is determined using data from a multimodal database, with atomic emission spectra as input and material properties and multimodal metadata as output.

[0096] The training method adopts a progressive training mode, consisting of two stages: pre-training and joint training. Pre-training uses an FNN network to extract shared features from the spectral data, reconstructs the spectrum using the FNN network, and uses MSEloss as the loss function for training, obtaining shared features representing the essence of the metallic material. Joint training uses the shared features to simultaneously train multimodal metadata, material properties, and content. Each training module uses an FNN network. Surface state, processing technology, and metal grade are used for classification model training, employing one-hot encoding and CrossEntropyLoss as the loss function. Excitation parameters, geometric state, and elemental content are used for regression model training, employing MSEloss as the loss function. The total loss function of the multimodal coherence model is:

[0097] Total loss = 0.1 × loss (surface state) + 0.1 × loss (morphology) + 0.6 × loss (metal grade) + 0.2 × loss (excitation parameters, geometric parameters, elemental composition)

[0098] Multimodal metadata, material properties, and elemental content of a material can be determined by analyzing its atomic emission spectrum.

[0099] III. Design of Self-Consistent Screening Rules:

[0100] The model-predicted multimodal information and material properties are used to search for matching elemental contents in a multimodal database. These are then compared with the model-predicted elemental contents. If the overall content deviation is ≤2%, the predicted content is corrected. If 2% < overall content deviation ≤5%, multimodal metadata and material properties matching the metal grade and composition are retrieved from the multimodal database. Simultaneously, the cosine similarity of the spectra is calculated to find the most similar corrected multimodal metadata and material properties. If no match is found, manual verification is triggered. If the overall content deviation >5%, historical data deviations are statistically analyzed to provide a confidence level, triggering manual verification.

[0101] IV. Output of Classification and Identification Results of the Metallic Materials to be Tested:

[0102] The atomic emission spectra of the tested metallic materials were obtained by using a LIBS instrument. After multimodal coherence model prediction and self-consistent screening, the classification and identification results are shown in Table 1.

[0103] Table 1 Classification and Recognition Results

[0104]

[0105]

[0106] Table 1 shows that the multimodal coherent model achieved good results in predicting grade, surface state, surface morphology, and elemental content. This demonstrates that the feature extraction layer extracted shared features representative of the material and environment. Combined with intelligent self-consistent classification validation, the prediction accuracy and confidence level were significantly improved.

[0107] Example 2

[0108] This invention also provides an intelligent self-consistent screening system for metallic materials based on a multimodal coherence model, applied to the aforementioned intelligent self-consistent screening method for metallic materials based on a multimodal database, comprising:

[0109] The multimodal database construction module collects atomic emission spectrum data of metallic materials under different morphologies and surface states, and associates them with material properties, acquisition parameters, and environmental parameters to generate a multimodal database; wherein, the morphology, surface state, acquisition parameters, and environmental parameters constitute multimodal metadata;

[0110] The multimodal coherent model building module is used to construct a feature extraction network using a deep learning network, perform cross-modal feature fusion on data in a multimodal database, extract shared features that reflect the essential properties of materials, train multiple sub-task models based on the shared features, and obtain a multimodal coherent model through weighted joint optimization. The multiple sub-task models include a material composition content prediction model, a material property classification model, and a multimodal metadata prediction model.

[0111] The self-consistent classification verification module is used to perform self-consistent verification and correct the results of the model output component content prediction results and material property classification results. The self-consistent verification methods include material property-content cross-verification, multimodal consistency check, and historical data backtesting evaluation.

[0112] Furthermore, embodiments of the present invention provide a computer-readable storage medium storing executable instructions, which, when executed, cause a processor to perform the intelligent self-consistent screening method for metallic materials based on a multimodal database as described in Embodiment 1.

[0113] Matters not covered in this invention are common knowledge.

[0114] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A smart self-consistent screening method for metallic materials based on a multimodal database, characterized in that, Includes the following steps: S1, Construct a multimodal database: Collect atomic emission spectral data of metallic materials under different morphologies and surface states, and associate them with material properties, acquisition parameters, and environmental parameters to generate a multimodal database; wherein, the morphology, surface state, acquisition parameters, and environmental parameters constitute multimodal metadata; S2, Establish a multimodal coherent model: A feature extraction network is constructed using a deep learning network to perform cross-modal feature fusion on data in the multimodal database, extract shared features that reflect the essential properties of materials, train multiple sub-task models based on the shared features, and obtain a multimodal coherent model through weighted joint optimization. The multiple sub-task models include a material composition content prediction model, a material property classification model, and a multimodal metadata prediction model. S3, Self-consistent classification verification: Perform self-consistent verification and correct the results of the model output component content prediction results and material property classification results. The self-consistent verification methods include material property-content cross-verification, multimodal consistency check, and historical data backtesting evaluation.

2. The intelligent self-consistent screening method for metallic materials based on a multimodal database according to claim 1, characterized in that, In S1, different forms include powder, block, thin film and coating; different surface states include oxide layer, corrosion layer and polished surface; atomic emission spectral data include full-spectrum wavelength, intensity, signal-to-noise ratio and spectral line broadening parameters; material properties include metal grade, chemical composition content, heat treatment state and processing technology; acquisition parameters include laser energy, laser pulse width, laser wavelength, number of ablation times, detection height, detection angle and spot position; environmental parameters include temperature and air pressure.

3. The intelligent self-consistent screening method for metallic materials based on a multimodal database according to claim 2, characterized in that, The multimodal database also includes: annotation information and knowledge graph, wherein the annotation information includes manually verified material property-content correspondence, abnormal spectrum markers, reference spectrum calibration, and cross-modal data association tags, and the knowledge graph includes material multimodal mapping relationships.

4. The intelligent self-consistent screening method for metallic materials based on a multimodal database according to claim 3, characterized in that, The multimodal database supports the following functions: Dynamic updates: By incorporating newly acquired multimodal metadata and validation results through an incremental learning mechanism, the material properties - content statistical distribution are automatically updated; Cross-modal retrieval: Quickly locate similar spectra based on morphology, surface condition, or acquisition parameters, and return historical data on corresponding material properties and content; Data traceability: Record the acquisition equipment, operators, and environmental conditions for each spectrum to ensure the results are reproducible; Semantic query: Supports retrieving relevant spectral data and model training results by material properties or component content range.

5. The intelligent self-consistent screening method for metallic materials based on a multimodal database according to claim 3, characterized in that, In S2, cross-modal feature fusion specifically includes: Data augmentation: Align and register atomic emission spectral data of the same metallic material under different modal metadata to construct intermodal correlations; Feature decoupling: Extracting shared features that reflect the essential properties of materials through a shared feature layer design; Dynamic weight allocation: Adaptively adjusts the contribution of different modal metadata to feature extraction based on morphology or surface state.

6. The intelligent self-consistent screening method for metallic materials based on a multimodal database according to claim 1, characterized in that, In S2, the training method includes: Independent training mode: The feature extraction network is optimized through unsupervised or self-supervised learning, and then the shared features obtained from pre-training are input into multiple sub-task models; Joint training mode: The feature extraction network, content prediction branch, and material property classification branch are constructed into an end-to-end model, which is simultaneously optimized through a multi-task loss function; Progressive training mode: First train the feature extraction network and a single subtask model, then introduce another subtask model for fine-tuning.

7. The intelligent self-consistent screening method for metallic materials based on a multimodal database according to claim 1, characterized in that, In S3, the material property-content cross-verification is: based on the standard component range corresponding to the material property, verify the rationality of the component content prediction results; Multimodal consistency check involves comparing the prediction results of the same metallic material under different acquisition parameters to eliminate random errors. Historical data retrospective evaluation involves calling historical data under the same modality metadata configuration in the multimodal database to conduct confidence assessment.

8. The intelligent self-consistent screening method for metallic materials based on a multimodal database according to claim 7, characterized in that, The material property-content cross-verification includes: Material property-driven validation: Based on the material property classification results output by the model, the standard content range corresponding to the material property is extracted from the multimodal database. If the predicted content of the component exceeds the range, the content correction is triggered. Material content-driven verification: Based on the composition content prediction results, the matching material properties in the multimodal database are retrieved in reverse. If the current material property is not in the set, manual verification is initiated. Dynamic threshold adjustment: Based on historical deviation data of similar metallic materials in a multimodal database, the material property-content compatibility threshold is adaptively adjusted.

9. A smart self-consistent screening system for metallic materials based on a multimodal coherence model, applied to the smart self-consistent screening method for metallic materials based on a multimodal database as described in claim 1, characterized in that, include: The multimodal database construction module collects atomic emission spectrum data of metallic materials under different morphologies and surface states, and associates them with material properties, acquisition parameters, and environmental parameters to generate a multimodal database; wherein, the morphology, surface state, acquisition parameters, and environmental parameters constitute multimodal metadata; The multimodal coherent model building module is used to construct a feature extraction network using a deep learning network, perform cross-modal feature fusion on data in a multimodal database, extract shared features that reflect the essential properties of materials, train multiple sub-task models based on the shared features, and obtain a multimodal coherent model through weighted joint optimization. The multiple sub-task models include a material composition content prediction model, a material property classification model, and a multimodal metadata prediction model. The self-consistent classification verification module is used to perform self-consistent verification and correct the results of the model output component content prediction results and material property classification results. The self-consistent verification methods include material property-content cross-verification, multimodal consistency check, and historical data backtesting evaluation.